Efficient Test-Time Adaptation For Robot Policies
In briefTest-time adaptation method (TEMPO) that helps robot walking policies adjust to changing conditions during deployment.
SummaryThe paper addresses reinforcement-learning locomotion policies for robots that degrade when conditions shift, such as rough terrain, payload changes, or sensing problems, and where retraining is impractical. It studies test-time adaptation, updating a pretrained policy online in a self-supervised way without rewards. The proposed TEMPO method adds an entropy regularizer on the critic and a score-based memory with importance sampling to stabilize updates under severe shift. Experiments on Go1, Go2, and G1 in MuJoCo Playground and on a physical Unitree Go2 show consistent gains, up to 80% in a single-device setting.